+/* Common tail code for NGG primitive shaders. */
+static void gfx10_emit_shader_ngg_tail(struct si_context *sctx,
+ struct si_shader *shader,
+ unsigned initial_cdw)
+{
+ radeon_opt_set_context_reg(sctx, R_0287FC_GE_MAX_OUTPUT_PER_SUBGROUP,
+ SI_TRACKED_GE_MAX_OUTPUT_PER_SUBGROUP,
+ shader->ctx_reg.ngg.ge_max_output_per_subgroup);
+ radeon_opt_set_context_reg(sctx, R_028B4C_GE_NGG_SUBGRP_CNTL,
+ SI_TRACKED_GE_NGG_SUBGRP_CNTL,
+ shader->ctx_reg.ngg.ge_ngg_subgrp_cntl);
+ radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN,
+ SI_TRACKED_VGT_PRIMITIVEID_EN,
+ shader->ctx_reg.ngg.vgt_primitiveid_en);
+ radeon_opt_set_context_reg(sctx, R_028A44_VGT_GS_ONCHIP_CNTL,
+ SI_TRACKED_VGT_GS_ONCHIP_CNTL,
+ shader->ctx_reg.ngg.vgt_gs_onchip_cntl);
+ radeon_opt_set_context_reg(sctx, R_028B90_VGT_GS_INSTANCE_CNT,
+ SI_TRACKED_VGT_GS_INSTANCE_CNT,
+ shader->ctx_reg.ngg.vgt_gs_instance_cnt);
+ radeon_opt_set_context_reg(sctx, R_028AAC_VGT_ESGS_RING_ITEMSIZE,
+ SI_TRACKED_VGT_ESGS_RING_ITEMSIZE,
+ shader->ctx_reg.ngg.vgt_esgs_ring_itemsize);
+ radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF,
+ SI_TRACKED_VGT_REUSE_OFF,
+ shader->ctx_reg.ngg.vgt_reuse_off);
+ radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
+ SI_TRACKED_SPI_VS_OUT_CONFIG,
+ shader->ctx_reg.ngg.spi_vs_out_config);
+ radeon_opt_set_context_reg2(sctx, R_028708_SPI_SHADER_IDX_FORMAT,
+ SI_TRACKED_SPI_SHADER_IDX_FORMAT,
+ shader->ctx_reg.ngg.spi_shader_idx_format,
+ shader->ctx_reg.ngg.spi_shader_pos_format);
+ radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL,
+ SI_TRACKED_PA_CL_VTE_CNTL,
+ shader->ctx_reg.ngg.pa_cl_vte_cntl);
+ radeon_opt_set_context_reg(sctx, R_028838_PA_CL_NGG_CNTL,
+ SI_TRACKED_PA_CL_NGG_CNTL,
+ shader->ctx_reg.ngg.pa_cl_ngg_cntl);
+
+ if (initial_cdw != sctx->gfx_cs->current.cdw)
+ sctx->context_roll = true;
+
+ if (shader->ge_cntl != sctx->last_multi_vgt_param) {
+ radeon_set_uconfig_reg(sctx->gfx_cs, R_03096C_GE_CNTL, shader->ge_cntl);
+ sctx->last_multi_vgt_param = shader->ge_cntl;
+ }
+}
+
+static void gfx10_emit_shader_ngg_notess_nogs(struct si_context *sctx)
+{
+ struct si_shader *shader = sctx->queued.named.gs->shader;
+ unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+ if (!shader)
+ return;
+
+ gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
+}
+
+static void gfx10_emit_shader_ngg_tess_nogs(struct si_context *sctx)
+{
+ struct si_shader *shader = sctx->queued.named.gs->shader;
+ unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+ if (!shader)
+ return;
+
+ radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
+ SI_TRACKED_VGT_TF_PARAM,
+ shader->vgt_tf_param);
+
+ gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
+}
+
+static void gfx10_emit_shader_ngg_notess_gs(struct si_context *sctx)
+{
+ struct si_shader *shader = sctx->queued.named.gs->shader;
+ unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+ if (!shader)
+ return;
+
+ radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
+ SI_TRACKED_VGT_GS_MAX_VERT_OUT,
+ shader->ctx_reg.ngg.vgt_gs_max_vert_out);
+
+ gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
+}
+
+static void gfx10_emit_shader_ngg_tess_gs(struct si_context *sctx)
+{
+ struct si_shader *shader = sctx->queued.named.gs->shader;
+ unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+ if (!shader)
+ return;
+
+ radeon_opt_set_context_reg(sctx, R_028B38_VGT_GS_MAX_VERT_OUT,
+ SI_TRACKED_VGT_GS_MAX_VERT_OUT,
+ shader->ctx_reg.ngg.vgt_gs_max_vert_out);
+ radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
+ SI_TRACKED_VGT_TF_PARAM,
+ shader->vgt_tf_param);
+
+ gfx10_emit_shader_ngg_tail(sctx, shader, initial_cdw);
+}
+
+unsigned si_get_input_prim(const struct si_shader_selector *gs)
+{
+ if (gs->type == PIPE_SHADER_GEOMETRY)
+ return gs->info.properties[TGSI_PROPERTY_GS_INPUT_PRIM];
+
+ if (gs->type == PIPE_SHADER_TESS_EVAL) {
+ if (gs->info.properties[TGSI_PROPERTY_TES_POINT_MODE])
+ return PIPE_PRIM_POINTS;
+ if (gs->info.properties[TGSI_PROPERTY_TES_PRIM_MODE] == PIPE_PRIM_LINES)
+ return PIPE_PRIM_LINES;
+ return PIPE_PRIM_TRIANGLES;
+ }
+
+ /* TODO: Set this correctly if the primitive type is set in the shader key. */
+ return PIPE_PRIM_TRIANGLES;
+}
+
+/**
+ * Prepare the PM4 image for \p shader, which will run as a merged ESGS shader
+ * in NGG mode.
+ */
+static void gfx10_shader_ngg(struct si_screen *sscreen, struct si_shader *shader)
+{
+ const struct si_shader_selector *gs_sel = shader->selector;
+ const struct tgsi_shader_info *gs_info = &gs_sel->info;
+ enum pipe_shader_type gs_type = shader->selector->type;
+ const struct si_shader_selector *es_sel =
+ shader->previous_stage_sel ? shader->previous_stage_sel : shader->selector;
+ const struct tgsi_shader_info *es_info = &es_sel->info;
+ enum pipe_shader_type es_type = es_sel->type;
+ unsigned num_user_sgprs;
+ unsigned nparams, es_vgpr_comp_cnt, gs_vgpr_comp_cnt;
+ uint64_t va;
+ unsigned window_space =
+ gs_info->properties[TGSI_PROPERTY_VS_WINDOW_SPACE_POSITION];
+ bool es_enable_prim_id = shader->key.mono.u.vs_export_prim_id || es_info->uses_primid;
+ unsigned gs_num_invocations = MAX2(gs_sel->gs_num_invocations, 1);
+ unsigned input_prim = si_get_input_prim(gs_sel);
+ bool break_wave_at_eoi = false;
+ struct si_pm4_state *pm4 = si_get_shader_pm4_state(shader);
+ if (!pm4)
+ return;
+
+ if (es_type == PIPE_SHADER_TESS_EVAL) {
+ pm4->atom.emit = gs_type == PIPE_SHADER_GEOMETRY ? gfx10_emit_shader_ngg_tess_gs
+ : gfx10_emit_shader_ngg_tess_nogs;
+ } else {
+ pm4->atom.emit = gs_type == PIPE_SHADER_GEOMETRY ? gfx10_emit_shader_ngg_notess_gs
+ : gfx10_emit_shader_ngg_notess_nogs;
+ }
+
+ va = shader->bo->gpu_address;
+ si_pm4_add_bo(pm4, shader->bo, RADEON_USAGE_READ, RADEON_PRIO_SHADER_BINARY);
+
+ if (es_type == PIPE_SHADER_VERTEX) {
+ /* VGPR5-8: (VertexID, UserVGPR0, UserVGPR1, UserVGPR2 / InstanceID) */
+ es_vgpr_comp_cnt = shader->info.uses_instanceid ? 3 : 0;
+
+ if (es_info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS]) {
+ num_user_sgprs = SI_SGPR_VS_BLIT_DATA +
+ es_info->properties[TGSI_PROPERTY_VS_BLIT_SGPRS];
+ } else {
+ num_user_sgprs = si_get_num_vs_user_sgprs(GFX9_VSGS_NUM_USER_SGPR);
+ }
+ } else {
+ assert(es_type == PIPE_SHADER_TESS_EVAL);
+ es_vgpr_comp_cnt = es_enable_prim_id ? 3 : 2;
+ num_user_sgprs = GFX9_TESGS_NUM_USER_SGPR;
+
+ if (es_enable_prim_id || gs_info->uses_primid)
+ break_wave_at_eoi = true;
+ }
+
+ /* If offsets 4, 5 are used, GS_VGPR_COMP_CNT is ignored and
+ * VGPR[0:4] are always loaded.
+ *
+ * Vertex shaders always need to load VGPR3, because they need to
+ * pass edge flags for decomposed primitives (such as quads) to the PA
+ * for the GL_LINE polygon mode to skip rendering lines on inner edges.
+ */
+ if (gs_info->uses_invocationid || gs_type == PIPE_SHADER_VERTEX)
+ gs_vgpr_comp_cnt = 3; /* VGPR3 contains InvocationID, edge flags. */
+ else if (gs_info->uses_primid)
+ gs_vgpr_comp_cnt = 2; /* VGPR2 contains PrimitiveID. */
+ else if (input_prim >= PIPE_PRIM_TRIANGLES)
+ gs_vgpr_comp_cnt = 1; /* VGPR1 contains offsets 2, 3 */
+ else
+ gs_vgpr_comp_cnt = 0; /* VGPR0 contains offsets 0, 1 */
+
+ si_pm4_set_reg(pm4, R_00B320_SPI_SHADER_PGM_LO_ES, va >> 8);
+ si_pm4_set_reg(pm4, R_00B324_SPI_SHADER_PGM_HI_ES, va >> 40);
+ si_pm4_set_reg(pm4, R_00B228_SPI_SHADER_PGM_RSRC1_GS,
+ S_00B228_VGPRS((shader->config.num_vgprs - 1) / 4) |
+ S_00B228_FLOAT_MODE(shader->config.float_mode) |
+ S_00B228_DX10_CLAMP(1) |
+ S_00B228_MEM_ORDERED(1) |
+ S_00B228_GS_VGPR_COMP_CNT(gs_vgpr_comp_cnt));
+ si_pm4_set_reg(pm4, R_00B22C_SPI_SHADER_PGM_RSRC2_GS,
+ S_00B22C_SCRATCH_EN(shader->config.scratch_bytes_per_wave > 0) |
+ S_00B22C_USER_SGPR(num_user_sgprs) |
+ S_00B22C_ES_VGPR_COMP_CNT(es_vgpr_comp_cnt) |
+ S_00B22C_USER_SGPR_MSB_GFX10(num_user_sgprs >> 5) |
+ S_00B22C_OC_LDS_EN(es_type == PIPE_SHADER_TESS_EVAL) |
+ S_00B22C_LDS_SIZE(shader->config.lds_size));
+
+ /* TODO: Use NO_PC_EXPORT when applicable. */
+ nparams = MAX2(shader->info.nr_param_exports, 1);
+ shader->ctx_reg.ngg.spi_vs_out_config =
+ S_0286C4_VS_EXPORT_COUNT(nparams - 1);
+
+ shader->ctx_reg.ngg.spi_shader_idx_format =
+ S_028708_IDX0_EXPORT_FORMAT(V_028708_SPI_SHADER_1COMP);
+ shader->ctx_reg.ngg.spi_shader_pos_format =
+ S_02870C_POS0_EXPORT_FORMAT(V_02870C_SPI_SHADER_4COMP) |
+ S_02870C_POS1_EXPORT_FORMAT(shader->info.nr_pos_exports > 1 ?
+ V_02870C_SPI_SHADER_4COMP :
+ V_02870C_SPI_SHADER_NONE) |
+ S_02870C_POS2_EXPORT_FORMAT(shader->info.nr_pos_exports > 2 ?
+ V_02870C_SPI_SHADER_4COMP :
+ V_02870C_SPI_SHADER_NONE) |
+ S_02870C_POS3_EXPORT_FORMAT(shader->info.nr_pos_exports > 3 ?
+ V_02870C_SPI_SHADER_4COMP :
+ V_02870C_SPI_SHADER_NONE);
+
+ shader->ctx_reg.ngg.vgt_primitiveid_en =
+ S_028A84_PRIMITIVEID_EN(es_enable_prim_id) |
+ S_028A84_NGG_DISABLE_PROVOK_REUSE(es_enable_prim_id);
+
+ if (gs_type == PIPE_SHADER_GEOMETRY) {
+ shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = es_sel->esgs_itemsize / 4;
+ shader->ctx_reg.ngg.vgt_gs_max_vert_out = gs_sel->gs_max_out_vertices;
+ } else {
+ shader->ctx_reg.ngg.vgt_esgs_ring_itemsize = 1;
+ }
+
+ if (es_type == PIPE_SHADER_TESS_EVAL)
+ si_set_tesseval_regs(sscreen, es_sel, pm4);
+
+ shader->ctx_reg.ngg.vgt_gs_onchip_cntl =
+ S_028A44_ES_VERTS_PER_SUBGRP(shader->ngg.hw_max_esverts) |
+ S_028A44_GS_PRIMS_PER_SUBGRP(shader->ngg.max_gsprims) |
+ S_028A44_GS_INST_PRIMS_IN_SUBGRP(shader->ngg.max_gsprims * gs_num_invocations);
+ shader->ctx_reg.ngg.ge_max_output_per_subgroup =
+ S_0287FC_MAX_VERTS_PER_SUBGROUP(shader->ngg.max_out_verts);
+ shader->ctx_reg.ngg.ge_ngg_subgrp_cntl =
+ S_028B4C_PRIM_AMP_FACTOR(shader->ngg.prim_amp_factor) |
+ S_028B4C_THDS_PER_SUBGRP(0); /* for fast launch */
+ shader->ctx_reg.ngg.vgt_gs_instance_cnt =
+ S_028B90_CNT(gs_num_invocations) |
+ S_028B90_ENABLE(gs_num_invocations > 1) |
+ S_028B90_EN_MAX_VERT_OUT_PER_GS_INSTANCE(
+ shader->ngg.max_vert_out_per_gs_instance);
+
+ /* User edge flags are set by the pos exports. If user edge flags are
+ * not used, we must use hw-generated edge flags and pass them via
+ * the prim export to prevent drawing lines on internal edges of
+ * decomposed primitives (such as quads) with polygon mode = lines.
+ *
+ * TODO: We should combine hw-generated edge flags with user edge
+ * flags in the shader.
+ */
+ shader->ctx_reg.ngg.pa_cl_ngg_cntl =
+ S_028838_INDEX_BUF_EDGE_FLAG_ENA(gs_type == PIPE_SHADER_VERTEX &&
+ !gs_info->writes_edgeflag);
+
+ shader->ge_cntl =
+ S_03096C_PRIM_GRP_SIZE(shader->ngg.max_gsprims) |
+ S_03096C_VERT_GRP_SIZE(shader->ngg.hw_max_esverts) |
+ S_03096C_BREAK_WAVE_AT_EOI(break_wave_at_eoi);
+
+ if (window_space) {
+ shader->ctx_reg.ngg.pa_cl_vte_cntl =
+ S_028818_VTX_XY_FMT(1) | S_028818_VTX_Z_FMT(1);
+ } else {
+ shader->ctx_reg.ngg.pa_cl_vte_cntl =
+ S_028818_VTX_W0_FMT(1) |
+ S_028818_VPORT_X_SCALE_ENA(1) | S_028818_VPORT_X_OFFSET_ENA(1) |
+ S_028818_VPORT_Y_SCALE_ENA(1) | S_028818_VPORT_Y_OFFSET_ENA(1) |
+ S_028818_VPORT_Z_SCALE_ENA(1) | S_028818_VPORT_Z_OFFSET_ENA(1);
+ }
+
+ shader->ctx_reg.ngg.vgt_reuse_off =
+ S_028AB4_REUSE_OFF(sscreen->info.family == CHIP_NAVI10 &&
+ sscreen->info.chip_external_rev == 0x1 &&
+ es_type == PIPE_SHADER_TESS_EVAL);
+}
+
+static void si_emit_shader_vs(struct si_context *sctx)
+{
+ struct si_shader *shader = sctx->queued.named.vs->shader;
+ unsigned initial_cdw = sctx->gfx_cs->current.cdw;
+
+ if (!shader)
+ return;
+
+ radeon_opt_set_context_reg(sctx, R_028A40_VGT_GS_MODE,
+ SI_TRACKED_VGT_GS_MODE,
+ shader->ctx_reg.vs.vgt_gs_mode);
+ radeon_opt_set_context_reg(sctx, R_028A84_VGT_PRIMITIVEID_EN,
+ SI_TRACKED_VGT_PRIMITIVEID_EN,
+ shader->ctx_reg.vs.vgt_primitiveid_en);
+
+ if (sctx->chip_class <= GFX8) {
+ radeon_opt_set_context_reg(sctx, R_028AB4_VGT_REUSE_OFF,
+ SI_TRACKED_VGT_REUSE_OFF,
+ shader->ctx_reg.vs.vgt_reuse_off);
+ }
+
+ radeon_opt_set_context_reg(sctx, R_0286C4_SPI_VS_OUT_CONFIG,
+ SI_TRACKED_SPI_VS_OUT_CONFIG,
+ shader->ctx_reg.vs.spi_vs_out_config);
+
+ radeon_opt_set_context_reg(sctx, R_02870C_SPI_SHADER_POS_FORMAT,
+ SI_TRACKED_SPI_SHADER_POS_FORMAT,
+ shader->ctx_reg.vs.spi_shader_pos_format);
+
+ radeon_opt_set_context_reg(sctx, R_028818_PA_CL_VTE_CNTL,
+ SI_TRACKED_PA_CL_VTE_CNTL,
+ shader->ctx_reg.vs.pa_cl_vte_cntl);
+
+ if (shader->selector->type == PIPE_SHADER_TESS_EVAL)
+ radeon_opt_set_context_reg(sctx, R_028B6C_VGT_TF_PARAM,
+ SI_TRACKED_VGT_TF_PARAM,
+ shader->vgt_tf_param);
+
+ if (shader->vgt_vertex_reuse_block_cntl)
+ radeon_opt_set_context_reg(sctx, R_028C58_VGT_VERTEX_REUSE_BLOCK_CNTL,
+ SI_TRACKED_VGT_VERTEX_REUSE_BLOCK_CNTL,
+ shader->vgt_vertex_reuse_block_cntl);
+
+ if (initial_cdw != sctx->gfx_cs->current.cdw)
+ sctx->context_roll = true;
+}
+